search

bbitmaster / ale_python_interface

A Python Interface for the Arcade Learning Environment (Shared Object)
GNU General Public License v2.0
126 stars 32 forks source link

Sending the same action causes ALE to hang #5

Closed el3ment closed 8 years ago

el3ment commented 8 years ago

Although in this example below, I am sending action 0 -- it's true for any action - tested on Breakout and Zaxxon

import sys
from ale_python_interface import ALEInterface
import numpy as np

ale = ALEInterface()

ale.loadROM("Breakout.bin")
legal_actions = ale.getLegalActionSet()

for episode in range(10):
    total_reward = 0.0 
    while not ale.game_over():
        print "acting"
        reward = ale.act(0);
        total_reward += reward
    print("Episode " + str(episode) + " ended with score: " + str(total_reward))
    ale.reset_game()

In order to avoid the hanging -- I resorted to performing a random action every 20 (a rough maximum value discovered via trial and error) frames. This bug came up when my Q-learner degenerated to choosing a single action.

el3ment commented 8 years ago

As a small update -- it looks like in Breakout, you have to send a "1" action.

el3ment commented 8 years ago

https://github.com/mgbellemare/Arcade-Learning-Environment/issues/154